SkillOrbit

Inspiration

Online learning platforms usually show the same generic recommendations to everyone. We wanted something smarter — a system that actually watches how a learner browses, searches, and bookmarks courses, then turns that real behavior into a clear, personalized next learning step.

The SmartReco Build Challenge 2026 gave us the perfect opportunity to build a production-grade, fully grounded recommendation engine instead of another toy demo.

What it does

SkillOrbit converts a user's browsing behavior into a personalized learning path.

• Silently tracks page views, searches, dwell time, and bookmarks
• Builds a live interest profile
• Uses a 7-stage LangGraph agent + Qdrant semantic retrieval
• Generates recommendations that are 100% grounded in the real catalog (no hallucinations)
• Shows full pipeline traces so every recommendation is explainable
• Delivers paths via shareable links and weekly email digests

How we built it

We built a complete production stack:

• Frontend + Backend: FastAPI + Jinja2 (server-rendered)
• Database: Supabase (PostgreSQL) for users, events, profiles & recommendations
• Vector Store: Qdrant for semantic search and RAG
• AI Layer: LangGraph (7-stage agent) + Mesh API (LLM + embeddings)
• Email: Resend for instant + weekly digests
• Scheduler: APScheduler for automated weekly recommendations
• Deployment: Render

The agent pipeline runs:
analyze → retrieve → evaluate → moderate → generate → validate → persist

Every recommended course ID is validated against the live catalog before it is shown to the user.

Challenges we ran into

1. Making recommendations truly grounded — preventing the LLM from inventing course IDs was harder than expected. We solved it with strict post-generation validation against the SQL catalog.
2. Non-blocking behavioral tracking — capturing events without slowing down the user experience required careful use of sendBeacon and batched requests.
3. Smart triggering — deciding when to regenerate a path (cooldown + meaningful behavior change) so we don't waste LLM calls.
4. Observability — building a full /trace page so judges (and users) can see exactly why a recommendation was made.

Accomplishments that we're proud of

• 100% grounded recommendations (zero hallucinated catalog items)
• Full end-to-end observability with Mesh trace IDs and Qdrant scores
• Beautiful guided demo at /demo that requires zero setup
• Production-ready dual-write (Admin → SQL + Qdrant)
• Weekly proactive email digests

What we learned

• How to design a reliable multi-stage LangGraph agent
• The importance of grounding and validation in RAG systems
• Building real behavioral intelligence instead of simple collaborative filtering
• Shipping a complete, judge-friendly product (not just a notebook)

What's next

• Better interest decay and long-term profile evolution
• Multi-path career orbit recommendations
• Mobile-friendly progressive web app
• Deeper integration with learning platforms for progress tracking

Built With

• api
• apscheduler
• behavioral
• fastapi
• jinja
• langgraph
• llm
• mesh
• postgresql
• python
• qdrant
• rag
• render
• resend
• search
• semantic
• supabase
• tracking
• vector